Agent readiness for AEO: an audit from machine-readable to recommended

A website should not be declared “ready for AI” merely because an assistant mentions the brand once. Cloudflare’s current model separates two questions: can an agent access and use the site, and does an assistant recommend the business in relevant answers? That distinction creates a useful audit sequence. First verify access, discovery and machine-readable delivery. Then assess mentions, citations and competitive visibility. Finally connect those observations to on-site behaviour and qualified commercial outcomes.

Cloudflare describes Agent Readiness Diagnostics as a hostname-level technical check and its AEO view as a recommendation assessment. Diagnostics examines signals such as robots.txt, XML sitemaps, response headers, Markdown delivery and published metadata for authentication or tools. The AEO capability tests likely customer prompts against Anthropic’s Claude and OpenAI’s GPT. Cloudflare reports citation rate, prominence, mention rate and share of voice. These are visibility indicators, not proof of revenue.

CreatikLab’s operational conclusion is therefore conditional: improve technical access where it is genuinely blocked, but do not treat a readiness score as an AEO outcome. A responsible programme needs separate evidence for retrieval, representation, recommendation, site engagement and lead qualification.

Cloudflare says Diagnostics fetches the site as an agent would, checks whether access is permitted, looks for discoverable content and callable interfaces, and returns pass, fail or neutral results with an evidence trail. Checks are grouped into quick wins, technical groundwork and advanced integrations. The official examples include crawler-readable robots rules, an XML sitemap, clean Markdown, content-use signals, API catalogues, link headers, authentication instructions, OAuth discovery, MCP, A2A cards, skills indexes, Web Bot Auth and WebMCP.

The same official page identifies commerce protocols as an informational category rather than part of the readiness score. It also says the AEO view infers an industry and category from a site, then uses discovery, comparison and advice prompts to observe assistant responses. Citation rate measures the share of sampled answers citing the site; mention rate records brand naming; prominence considers how much cited material appears and how early; share of voice compares citations with competitors.

The announcement does not specify pricing, account eligibility, geographic rollout, prompt volume, testing frequency, model-version controls or a guaranteed relationship between its metrics and sales. Those omissions belong in the evaluation record rather than being filled with assumptions.

CreatikLab uses a four-state matrix so teams do not prescribe content when the failure is technical, or deploy infrastructure when the problem is weak market evidence. Classify a representative set of commercial pages before approving work.

• Readable and recommended: preserve access, inspect which claims earn citations, and test whether recommended journeys generate sales-accepted enquiries.
• Readable but not recommended: investigate proposition clarity, evidence, entity consistency and useful differentiation before creating more pages.
• Not reliably readable but already mentioned: protect existing brand demand while correcting crawl, discovery or machine-readable delivery defects.
• Neither readable nor recommended: repair foundational access first, then run a controlled content and prompt test instead of promising immediate visibility.

Each classification needs attached proof: the exact requested URL, response status, relevant headers, rendered or machine-readable content, the prompt wording, assistant and test date, returned citation, competing sources, landing session and CRM disposition where available. A screenshot alone is insufficient because it cannot establish how the page was retrieved or whether an enquiry became commercially valid.

An inspectable audit assigns every finding to an evidence object, a corrective action and a named owner. This prevents an AEO dashboard from becoming an unprioritised list of scores.

1. Access — Evidence: robots rules, response headers and fetched response. Action: remove accidental blocking without opening restricted areas. Owner: engineering and security.
2. Discovery — Evidence: sitemap inclusion and internal route to the page. Action: reconcile canonical commercial URLs and remove orphan paths. Owner: SEO and web engineering.
3. Readable content — Evidence: the clean text or Markdown an agent receives. Action: preserve prices, claims, exclusions and proof in unambiguous text where the business publishes them. Owner: content and product.
4. Entity consistency — Evidence: matching brand, service and organisation details across priority pages. Action: resolve contradictions at the authoritative record. Owner: brand and content operations.
5. Recommendation sample — Evidence: logged prompts, assistant responses, citations and competitors. Action: diagnose the missing evidence or unclear proposition rather than imitating an answer. Owner: AEO strategist.
6. Commercial outcome — Evidence: analytics events, consent state and CRM status. Action: connect visits and enquiries to sales acceptance. Owner: analytics, revenue operations and sales.

A finding is complete only when its reproduction steps, risk, owner and acceptance test are recorded. “Add schema” or “write more content” is not an actionable diagnosis without a demonstrated defect.

Begin with a bounded page set: core service pages, decisive comparison or eligibility pages, proof-rich documentation and the routes that convert demand. For each page, define the customer decision it should support. Then test whether an agent can discover the URL, extract the decisive information and distinguish the offer from alternatives.

Where a technical defect exists, correct it in the delivery layer and retest the same URL. Where information is vague, revise the page around an actual buyer question, verifiable claims, limitations and next actions. Where evidence is absent, obtain it from the accountable business owner; do not ask generative AI to manufacture experience, policies or results. Where several prompts express the same decision, strengthen one authoritative page instead of creating near-duplicates.

MCP, WebMCP, API catalogues or agent authentication may be relevant when an agent must perform an authorised task rather than merely read public information. CreatikLab treats those as architecture decisions requiring threat modelling, permission boundaries, logs and human approval. Cloudflare lists these interfaces as advanced checks, but its announcement does not establish that every marketing website needs them.

Measurement should retain four separate layers. Layer one is technical availability: successful retrieval, correct status, expected headers and usable machine-readable content. Layer two is sampled answer visibility: mention rate, citation rate, prominence and share of voice, using the definitions Cloudflare provides. Layer three is site response: attributable sessions, engaged visits and completion of meaningful actions. Layer four is commercial quality: enquiries accepted by sales because they match the agreed customer profile, need, territory or service availability.

Record the assistant, prompt family, exact wording, locale, observation time, cited URL and response evidence for every sampled answer. Compare like with like; do not merge recommendation prompts with informational prompts or one language with another. Assistant responses can vary, so report distributions and repeated observations rather than presenting one answer as a stable ranking.

For qualified-lead reporting, define the CRM statuses before testing. At minimum distinguish raw enquiry, reachable contact, relevant need, accepted opportunity and disqualified reason. Cloudflare’s announced AEO metrics do not provide this commercial classification. It must come from consent-aware analytics and the company’s sales process. The objective is to discover useful relationships, not to promise lead volume.

Opening access indiscriminately can expose private routes, duplicate environments or content that was never approved for public use. Changes to robots rules, authentication metadata or callable interfaces therefore require security review. Public readability and permission to execute an action are different states. An agent that can read a service description should not automatically gain access to customer data, account changes or transactions.

Recommendation testing also carries interpretation risk. A sampled prompt set can overrepresent the language chosen by the evaluator. An inferred category may not match the company’s actual commercial segment. Citation rate may rise while the brand is described inaccurately, and mention rate may rise without a link. These cases require response-level review, not celebration of an aggregate score.

Set rollback criteria before implementation. Preserve previous headers and robots configurations, monitor unintended crawler changes, and recheck critical conversion routes after deployment. Keep human approval for claims, access policy and production release. Automation can collect evidence and repeat checks; accountability for publication and access remains with named people.

• Do not assume a perfect readiness result guarantees inclusion, citation, recommendation, traffic or sales.
• Do not assume every AI crawler should receive identical access; content policy and security requirements may differ.
• Do not assume Markdown replaces accessible, useful HTML for human visitors.
• Do not assume a mention is a citation, or that either represents a qualified lead.
• Do not assume Cloudflare’s sampled assistants represent every answer engine, locale or future model version.
• Do not assume inferred competitors or categories match the commercial set used by sales.
• Do not assume advanced agent interfaces are necessary when the task is only public content discovery.
• Do not assume the announcement confirms pricing, plan access, rollout coverage or testing cadence; it does not specify them.

Provider comparisons should therefore ask for inspectable deliverables: a URL-level access inventory, response evidence, a documented prompt protocol, a prioritised remediation register, release acceptance tests, analytics and CRM mapping, security ownership and a clear explanation of uncertainty. A provider should be able to show why each intervention follows from a recorded defect.

The practical first engagement is an agent-readiness and AEO evidence audit, not a promise of recommendations. Deliverables should include a representative URL inventory, crawl and machine-readable response captures, robots and sitemap review, content-extraction samples, a controlled prompt matrix, mention-versus-citation analysis, remediation priorities, owners, acceptance tests and a measurement map from observed visits to CRM-qualified outcomes.

Explore CreatikLab’s AI automation service when the diagnosis may require repeatable checks, custom integrations or governed agent interfaces. Use the AI Expert route to bridge the findings into an architecture and operating decision with senior review. These routes do not imply that every site needs an agent integration; the audit determines whether content, infrastructure, measurement or governance is the actual constraint.

If the situation is not yet clear, tell Lia about the affected site, priority journeys and available evidence. Include the public hostname, important services, known access controls, target markets, analytics setup and how sales currently defines an acceptable enquiry. That context allows diagnosis to continue without substituting a generic score for the business problem.

What is an agent-readiness audit?

It is a documented check of whether agents can access, discover and interpret the intended public website content, followed by acceptance tests and ownership for each defect. It should not be treated as proof that assistants will recommend the business.

Which Cloudflare AEO metrics are confirmed?

Cloudflare describes citation rate, prominence, mention rate and share of voice. These measure sampled recommendation visibility in different ways; none is presented as a qualified-lead metric.

Does Cloudflare test every AI assistant?

No such coverage is stated. The announcement names Anthropic’s Claude and OpenAI’s GPT. It does not specify support for every engine, locale or model version.

Does a clean robots.txt guarantee an AI citation?

No. Crawler access can remove a technical obstacle, but recommendation also depends on the response context and available information. Cloudflare does not promise citations from a passed technical check.

Should every website publish MCP or WebMCP interfaces?

No. Cloudflare lists agent-native interfaces among advanced readiness checks, but the appropriate architecture depends on whether an authorised agent action is genuinely required. Public information retrieval alone may not justify a callable interface.

How should qualified leads be measured?

Connect consent-aware site events to CRM stages such as relevant need, reachable contact, sales acceptance and disqualification reason. Keep that commercial measurement separate from mentions, citations and share of voice.